EXIN BCS Gen AI Award key terms

A class of machine learning systems that produce new content such as text, images, audio, or code by learning statistical patterns from large training datasets, rather than only classifying or predicting labels.

A neural network trained on very large text corpora to predict and generate language, typically built on the transformer architecture and capable of summarisation, translation, question answering, and code generation.

A broad, general-purpose model trained on massive unlabelled data that can be adapted to many downstream tasks through prompting or fine-tuning instead of being built for one narrow purpose.

A neural network architecture based on self-attention that processes all parts of an input in parallel, enabling the scale and context handling behind modern language and multimodal models.

A unit of text, roughly a word fragment, that a model reads and generates; token counts determine context limits, latency, and API cost.

A numeric vector representing the meaning of text, images, or other data so that semantically similar items sit close together, enabling search, clustering, and retrieval.

The collection of examples a model learns from; its quality, coverage, recency, and licensing strongly shape model behaviour, bias, and legal exposure.

Further training of a pre-trained model on a smaller, task-specific dataset so it adapts to a domain, tone, or format while retaining its general capabilities.

The act of running a trained model to produce an output for a given input; inference cost and speed drive operational decisions in deployed generative AI services.

The learned numeric weights inside a neural network that encode its behaviour; parameter count is a rough indicator of capability, cost, and hardware requirements.

A model that accepts or produces more than one data type, such as text plus images or audio, allowing tasks like describing a photo or generating an image from a written prompt.

A generative technique that learns to reverse a gradual noising process, producing images or audio by starting from random noise and iteratively refining it toward a coherent result.

A confident but incorrect or fabricated model output, such as an invented citation or non-existent policy, arising because the system predicts plausible text rather than verifying facts.

The maximum amount of text, measured in tokens, that a model can consider at once; content beyond it is truncated or ignored, affecting long-document and chat use cases.

The date after which a model's training data contains no information, meaning recent events, prices, or regulations are unknown unless supplied in the prompt or retrieved externally.

Anchoring model output in verifiable sources such as supplied documents, databases, or search results so that responses reflect evidence rather than unsupported generation.

An architecture that retrieves relevant passages from an external knowledge store and inserts them into the prompt so the model answers from current, organisation-specific evidence.

A data store optimised for similarity search over embeddings, used to find the passages most relevant to a user query before they are passed to a generative model.

The instruction and any supporting context given to a generative AI system that steers what it produces; its wording, structure, and examples strongly influence output quality.

The practice of designing, testing, and iterating instructions and context to obtain reliable, relevant, and appropriately formatted outputs from a generative AI model.

A persistent instruction set that defines a model's role, constraints, tone, and boundaries before user input arrives, shaping behaviour across an entire conversation or application.

Asking a model to perform a task with instructions only and no worked examples, relying on its general training to infer the expected behaviour and output format.

Including a small number of example inputs and desired outputs in the prompt so the model can imitate the pattern, tone, or format required for the real task.

A technique that asks a model to reason through intermediate steps before answering, improving performance on multi-step problems while increasing output length and cost.

Assigning the model a persona or expertise, such as auditor or support agent, to steer vocabulary, depth, and perspective in its response.

A sampling setting controlling randomness in generation; lower values give more deterministic, repeatable output, while higher values increase variety and creative divergence.

An attack in which malicious text, often hidden in documents or web pages, tricks a model into ignoring its instructions or leaking data or capabilities.

Unintended exposure of confidential information, occurring when sensitive input is sent to an external model, retained in logs, or reproduced in outputs visible to others.

Technical and procedural controls such as content filters, input validation, output checks, and usage policies that constrain what a generative AI system will accept or produce.

Deliberate human review of AI outputs before consequential action, ensuring accountability, catching errors, and satisfying regulatory expectations for meaningful human control.

Systematic skew in training data that causes outputs to under-represent or stereotype groups, producing unfair or inaccurate results in hiring, credit, healthcare, and similar decisions.

Openness about when and how generative AI is used, what data it processes, and what its limitations are, so users and affected parties can make informed judgements.

The degree to which the reasoning or contributing factors behind an AI output can be understood and communicated, which is difficult for large neural networks and matters in regulated decisions.

Clear assignment of responsibility for AI-assisted outcomes to identifiable people or organisations, so that errors, harms, and compliance failures have an owner.

The legal exposure arising when generated content reproduces protected material or when training data was used without permission, requiring licence checks and originality review.

Informing users or audiences that content was generated or assisted by AI, supporting honesty, trust, and compliance with emerging labelling obligations.

The policies, roles, risk assessments, and monitoring an organisation puts in place to ensure generative AI use is lawful, safe, and aligned with business objectives.

Structured evaluation of a proposed generative AI application against value, feasibility, data sensitivity, and risk criteria to decide whether and how to proceed.

The financial benefit of a generative AI initiative relative to its total cost, including licences, integration, training, review effort, and ongoing monitoring.

Using generative AI to handle repetitive knowledge work such as drafting, summarising, classifying, and extracting information, freeing staff for judgement-intensive tasks.